Predicting Psychotic Relapse Using Speech-Based Early Detection

Likelihood of relapse estimated using Speech-NLP Metrics

This primary outcome will assess the ability of speech-based NLP metrics (coherence, connectedness, and complexity) to predict impending relapses in psychosis. Monthly speech samples will be analyzed to determine if changes in these metrics can distinguish timepoints preceding relapses from those not followed by relapse, with the aim of predicting relapses up to four weeks in advance. The likelihood of relapse is a numerical probabilistic estimate without any units. Outcome definition: Occurrence of relapse (i.e., psychiatric hospitalization, an increase in the level of psychiatric care, or substantial clinical deterioration \>1wk that requires \>25% increase in Defined Daily Dose equivalents of antipsychotics)

Generalization of Speech-Based Relapse Prediction Models Across Languages and Genders

This outcome will assess whether the speech-based relapse prediction models are valid and perform equally well across different languages (English and French) and genders (male and female). The study will evaluate how sociodemographic factors such as language and sex impact the predictive accuracy of the models. NLP metrics (coherence, connectedness, complexity) will be correlated with clinical outcomes, and model performance will be compared across linguistic and gender subgroups to ensure generalizability.

Likelihood of relapse estimated using multi-level speech features

This exploratory outcome will assess the likelihood of relapse prediction by integrating acoustic features (e.g., speech rate, intonation), prosodic features (e.g., rhythm, pitch), and sentiment analysis into the primary measures (coherence, connectedness, complexity). Accuracy (the proportion of all relapse classifications that were correct, whether positive or negative) will be reported for the predictive models.